Diabetic Retinopathy (DR) is a consequence of diabetes mellitus that results in damage to the retina’s blood vessel networks. It is now the major cause of irreversible blindness among diabetics. It is more prevalent in those with a history of diabetes mellitus. A high blood glucose level results in the loss of blood and other fluids from the retina’s blood vessels. The importance of early detection and treatment of this illness cannot be overstated. In medical image processing, reliably detecting diabetic retinopathy from digital fundus pictures is considered an open challenge that requires the exploration of innovative approaches. An automated computerized system accomplishes this goal via a series of processes, including the identification and classification of lesions in fundus images. We are motivated by the researchers who chose to use deep learning for the detection of many diseases, like glaucoma, due to its recent breakthroughs and significant success over typical machine learning techniques in a variety of applications. The goal of this article is to develop reformed networks for detecting and categorizing diabetic retinopathy into five classes. Through this work, three novel Convolutional Neural Networks(CNNs) based on deep learning have been proposed for the prediction of diabetic retinopathy. The first proposed network is built from scratch, the second network is the ensemble of five best-performing networks, and the last one is the Convolutional Neural Network-Long Short Term Memory (CNN-LSTM) model. The performance of three proposed networks is compared with twenty-one well-accepted image nets. The effectiveness of the proposed reformed networks is validated using multiple performance indicators on seven experiments of three Kaggle datasets (separately and in combination of these). The presented work achieves the maximum accuracy of 0.9545 (highest average-wise 0.9368), a maximum F1 score of 0.9685 (highest average-wise 0.9374), the maximum sensitivity of 0.9566 (highest average-wise 0.9420),and a maximum AUC score of 0.9769 (highest average-wise 0.9395). The results of the experiments showed that the provided model outperforms the many current models in terms of classification. So, the results of the calculations show that with the proposed networks, retinal problems in diabetic patients can be effectively found so that they can get a better diagnosis and avoid losing their sight.